@zostaff: This paper completely changed how I think about session compaction: Model the DAG -> Snapshot -> Branch -> Trim structu…
Summary
A paper presents a structural approach to session compaction, modeling chat history as a DAG with snapshots and branches to trim bloat while preserving meaning, achieving a mean 20% token cut in coding sessions.
View Cached Full Text
Cached at: 08/05/26, 04:26 PM
This paper completely changed how I think about session compaction:
Model the DAG -> Snapshot -> Branch -> Trim structurally -> Keep verbatim
Here is the 5-step blueprint:
Model the DAG: Session history is not a linear log but version-controlled state, a directed acyclic graph with snapshots and branches.
Snapshot: Accumulated understanding, architectural maps, trade-offs, codebase conventions, is captured as named state instead of drowning at window overflow.
Branch: A snapshot is reused across independent parallel sessions, context forked with lineage tracking preserved.
Trim structurally: A three-pass algorithm strips mechanical bloat, raw tool outputs, base64 images, metadata, not meaning.
Keep verbatim: Every user message and assistant response stays word for word, the trimming is structurally lossless.
The key insight: Native autocompaction crushes 98 percent of a session’s accumulated state into a short summary, and every bit of understanding earned across the session is lost.
Across 76 real coding sessions this gave a mean 20 percent token cut and up to 86 on heavy ones, while mixed sessions break even within 10 turns under prompt caching.
Read this, then check the article below.
Similar Articles
@zostaff: This paper completely changed how I think about self-evolving codebases: Backlog -> Ideate -> Triage -> Execute -> Poli…
This paper presents a six-phase blueprint for self-evolving codebases using LLMs, emphasizing convergence toward a specification rather than metrics.
@omarsar0: // The Efficiency Frontier // Cool paper on context management. As agents reuse the same documents and histories across…
This paper introduces The Efficiency Frontier, a unified framework for cost–performance optimization in LLM context management that models context strategy selection as a deployment-aware optimization problem, achieving 25% reduction in token usage and over 50% lower token cost with amortized memory compression compared to full-context prompting.
@zostaff: This paper completely changed how I think about self-improving agents: Initialize -> Run -> Analyze -> Branch -> Update…
This paper presents a novel blueprint for self-improving agents that combines scaffold editing and weight training through a meta-agent and feedback-agent, achieving a 14x speedup on a CUDA kernel for AlphaFold.
My last 2 papers with time stamped logs in continuity and emerging personality in LLM based entities & how memory reduces tokens consumption and developers time
The author presents two papers examining time-stamped logs for continuity and emerging personality in LLM-based entities, and how memory reduces token consumption and developer time.
Beyond Compaction: Structured Context Eviction for Long-Horizon Agents
Introduces Context Window Lifecycle (CWL), a structured context eviction scheme for long-horizon LLM agents that maintains an effectively unbounded working horizon by evicting content based on a dependency graph, avoiding the limitations of summarization-based compaction and recency truncation.